Fault Detection Using Model-Based Approaches: A Case Study on a Fluid Mechanics Test Bench - Training Data and Model Weights

About This dataset is part of the publication Kruse et al., Fault Detection Using Model-Based Approaches: A Case Study on a Fluid Mechanics Test Bench. 2026. It includes raw wind tunnel data, synthetically generated datasets, and trained LSTM model weights (exported from KERAS). This dataset is licensed under CC BY 4.0. Publication abstract This study examines the development of a method for fault detection, applied to the calibration of a five-hole probe in an aerodynamic test bench. The method is based on model-based prediction of measurement values, with two different model types investigated for modelling the underlying system: a physics-based model, based on equations for a Greitzer Compression System, and a data-driven model, based on a long short-term memory (LSTM) neural network. The fault detection performance is evaluated using a novel benchmarking method by generating synthetic faults. The results show that the neural network is capable of predicting system behaviour more accurately, with a median deviation below 1\%, and detecting faults more reliably, with an $f_1$ score of 0.91. A hybrid approach, combining both methods, is also investigated and demonstrates comparable performance in predicting system behaviour compared to the data-driven model while requiring fewer training data. The effects of varying error detection thresholds are also investigated and indicate a lower sensitivity of error detection quality for the data-driven modelling approach. Data Content The published data is provided as a zip-file. It includes the following: data_real: real data, measured on a calibration wind tunnel, used for training, threshold determination and benchmarking; raw text data_synthetic: synthetically generated data using the physic-based model, used for training of the hybrid model; raw text trained_nns: complete neural networks training_data_variation: data-driven model trained on 1 to 7 (real) datasets, threshold factor theta = 1.05 threshold_variation: data-driven model trained on 7 (real) datasets, threshold factor theta varied from 1.02 to 1.50 hybrid_model: hybrid model trained on either 6 datasets (synthetic only) or 8 datasets (6 synthetic and 2 real), also includes a hybrid model trained on 8 datasets with manually adjusted threshold values *.info: raw text, includes information about hyperparameters, input and output features, normalization values *.keras: Keras model file, can be imported into Python following this guide *.thresh: raw text, includes threshold values for the output features for their respective fan speed bins